Create README.md
Browse files
README.md
ADDED
@@ -0,0 +1,77 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: mit
|
3 |
+
---
|
4 |
+
# TAC depth encoder
|
5 |
+
|
6 |
+
<!-- Provide a quick summary of what the model is/does. -->
|
7 |
+
|
8 |
+
This model is used for encoding a depth image into a dense feature.
|
9 |
+
|
10 |
+
## Model Details
|
11 |
+
|
12 |
+
### Model Description
|
13 |
+
|
14 |
+
<!-- Provide a longer summary of what this model is. -->
|
15 |
+
|
16 |
+
The model is pre-trained with RGB-D contrastive objectives, named TAC.
|
17 |
+
Different from InfoNCE-based loss fuctions, TAC leverages the similarity between videos frames and estimate a similarity matrix as soft labels.
|
18 |
+
The backbone of this version is ViT-B/32.
|
19 |
+
The pre-training is conducted on a new unified RGB-D database, UniRGBD.
|
20 |
+
|
21 |
+
### Model Sources
|
22 |
+
|
23 |
+
<!-- Provide the basic links for the model. -->
|
24 |
+
|
25 |
+
- **Repository:** [TAC](https://github.com/RavenKiller/TAC)
|
26 |
+
- **Paper:** [Learning Depth Representation from RGB-D Videos by Time-Aware Contrastive Pre-training](https://ieeexplore.ieee.org/document/10288539)
|
27 |
+
|
28 |
+
## Uses
|
29 |
+
|
30 |
+
<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
|
31 |
+
|
32 |
+
### Direct Uses
|
33 |
+
|
34 |
+
```
|
35 |
+
from transformers import CLIPImageProcessor, CLIPVisionModel, CLIPVisionConfig
|
36 |
+
import numpy as np
|
37 |
+
tac_depth_model = CLIPVisionModel.from_pretrained("RavenK/TAC-ViT-base")
|
38 |
+
tac_depth_processor = CLIPImageProcessor.from_pretrained("RavenK/TAC-ViT-base")
|
39 |
+
|
40 |
+
# Assume test.png is a depth image with a scale factor 1000
|
41 |
+
MIN_DEPTH = 0.0
|
42 |
+
MAX_DEPTH = 10.0
|
43 |
+
DEPTH_SCALE = 1000
|
44 |
+
|
45 |
+
depth_path = "test.png"
|
46 |
+
depth = Image.open(depth_path)
|
47 |
+
depth = np.array(depth).astype("float32") / DEPTH_SCALE # to meters
|
48 |
+
depth = np.clip(depth, MIN_DEPTH, MAX_DEPTH) # clip to [MIN_DEPTH, MAX_DEPTH]
|
49 |
+
depth = (depth - MIN_DEPTH) / (MAX_DEPTH - MIN_DEPTH) # normalize to [0,1]
|
50 |
+
depth = np.expand_dims(depth, axis=2).repeat(3, axis=2) # extend to 3 channels
|
51 |
+
depth = tac_depth_processor(depth, do_rescale=False, return_tensors="pt").pixel_values # preprocess (resize, normalize and to tensor)
|
52 |
+
|
53 |
+
outputs = tac_depth_model(pixel_values=depth)
|
54 |
+
outputs = outputs["last_hidden_state"][:, 0, :] # get embedding without FC. may be used for other downstream fine-tuning
|
55 |
+
```
|
56 |
+
|
57 |
+
### Other Uses
|
58 |
+
|
59 |
+
Please refer to our code repository to get more details.
|
60 |
+
|
61 |
+
## Citation
|
62 |
+
|
63 |
+
<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
|
64 |
+
|
65 |
+
```
|
66 |
+
@ARTICLE{10288539,
|
67 |
+
author={He, Zongtao and Wang, Liuyi and Dang, Ronghao and Li, Shu and Yan, Qingqing and Liu, Chengju and Chen, Qijun},
|
68 |
+
journal={IEEE Transactions on Circuits and Systems for Video Technology},
|
69 |
+
title={Learning Depth Representation from RGB-D Videos by Time-Aware Contrastive Pre-training},
|
70 |
+
year={2023},
|
71 |
+
volume={},
|
72 |
+
number={},
|
73 |
+
pages={1-1},
|
74 |
+
doi={10.1109/TCSVT.2023.3326373}}
|
75 |
+
```
|
76 |
+
|
77 |
+
|